AUV Navigation and Localization - A Review

Transcription

1 1 AUV Navigation and Localization - A Review Liam Paull, Sajad Saeedi, Mae Seto and Howard Li Abstract Autonomous underwater vehicle (AUV) navigation and localization in underwater environments is particularly challenging due to the rapid attenuation of GPS and radio frequency signals. Underwater communications are low bandwidth and unreliable and there is no access to a global positioning system. Past approaches to solve the AUV localization problem have employed expensive inertial sensors, used installed beacons in the region of interest, or required periodic surfacing of the AUV. While these methods are useful, their performance is fundamentally limited. Advances in underwater communications and the application of simultaneous localization and mapping (SLAM) technology to the underwater realm have yielded new possibilities in the field. This paper presents a review of the state of the art of underwater autonomous vehicle navigation and localization, as well as a description of some of the more commonly used methods. In addition, we highlight areas of future research potential. I. INTRODUCTION The development of autonomous underwater vehicles (AUVs) began in earnest in the 1970s. Since then, advancements in the efficiency, size, and memory capacity of computers have enhanced that potential. As a result, many tasks that were originally achieved with towed arrays or manned vehicles are being completely automated. AUV designs include torpedo-like, gliders, and hovering, and their sizes range from human-portable to hundreds of tons. AUVs are now being used for a variety of tasks, including oceanographic surveys, demining, and bathymetric data collection in marine and riverine environments. Accurate localization and navigation is essential to ensure the accuracy of the gathered data for these applications. A distinction should be made between navigation and localization. Navigational accuracy is the precision with which the AUV guides itself from one point to another. Localization accuracy is the error in how well the AUV localizes itself within a map. AUV navigation and localization is a challenging problem due primarily to the rapid attenuation of higher frequency signals and the unstructured nature of the undersea environment. Above water, most autonomous systems rely on radio or spread spectrum communications and global positioning. However, underwater such signals propagate only short distances and acoustic based sensors and communications perform better. Acoustic communications still suffer from many shortcomings such as: Small bandwidth, which means communicating nodes have had to use time division multiple access (TDMA) techniques to share information, Low data rate, which generally constrains the amount of data that can be transmitted, High latency since the speed of sound in water is only 1500m/s (slow compared with the speed of light), Variable sound speed due to fluctuating water temperature and salinity, Multi-path transmissions due to the presence of an upper (free surface) and lower (sea bottom) boundary coupled with highly variable sound speed Unreliability, resulting in the need for a communications system designed to handle frequent data loss in transmissions. Liam Paull is a COBRA team member with the Department of Electrical and Computer Engineering, University of New Brunswick, Fredericton, Canada, liam.paull@unb.ca Sajad Saeedi G. is a COBRA team member with the Department of Electrical and Computer Engineering, University of New Brunswick, Fredericton, Canada, sajad.saeedi.g@unb.ca Mae Seto is with Defence Research and Development Canada, Halifax, Nova Scotia, Canada, Mae.Seto@drdc-rddc.gc.ca Howard Li is a COBRA team member with the Department of Electrical and Computer Engineering, University of New Brunswick, Fredericton, Canada, howard@unb.ca

2 2 Notwithstanding these significant challenges, research in AUV navigation and localization has exploded in the last ten years. The field is in the midst of a paradigm shift from old technologies, such as long baseline (LBL) and ultra short baseline (USBL), which require pre-deployed and localized infrastructure, towards dynamic multiagent system approaches that allow for rapid deployment and flexibility with minimal infrastructure. In addition, simultaneous localization and mapping (SLAM) techniques developed for above ground robotics applications are being increasingly applied to underwater systems. The result is that bounded error and accurate navigation for AUVs is becoming possible with less cost and overhead. A. Outline AUV navigation and localization techniques can be categorized according to Fig. 1. This review paper will be organized based on this structure. AUV Navigation Sensor Technology inertial / dead reckoning acoustic transponders and modems Geophysical optical acoustic magnetic compass DVL IMU pressure sensor monocular Cameras stereo magnetometer single fixed beacon Transponders short baseline Sensor Hardware ultra-short baseline long baseline / GIBs acoustic modem imaging sonar SONAR ranging sonar manned surface support autonomous surface crafts homogeneous teams heterogeneous teams Collaboration Cooperative Localization Fig. 1. Outline of underwater navigation classifications. These methods are often combined in one system to provide increased performance. In general, these techniques fall into one of three main categories: Inertial / Dead Reckoning: Inertial navigation uses accelerometers and gyroscopes for increased accuracy to propagate the current state. Nevertheless, all of the methods in this category have position error growth that is unbounded. Acoustic transponders and modems: Techniques in this category are based on measuring the time-of-flight (TOF) of signals from acoustic beacons or modems to perform navigation. Geophysical: Techniques that use external environmental information as references for navigation. This must be done with sensors and processing that are capable of detecting, identifying, and classifying some environmental features. Sonar sensors are based on acoustic signals, however, navigation with imaging or bathymetric sonars are based on detection, identification, and classification of features in the environment. Therefore, navigation that is sonar-based

3 3 falls into both the acoustic and geophysical categories. A distinction is made between sonar and other acoustic based navigation schemes, which rely on externally generated acoustic signals emitted from beacons or other vehicles. The type of navigation system used is highly dependent on the type of operation or mission and that in many cases different systems can be combined to yield increased performance. The most important considerations are the size of the region of interest and the desired localization accuracy. Past reviews on this topic include [1], [2], and [3]. Significant advances have been made since these reviews both in previously established technologies, and in new areas. In particular, the development of acoustic communications through the use of underwater modems has lead to the development of new algorithms. In addition, advancements in SLAM research has been applied to the underwater domain in a number of new ways. II. BACKGROUND Most modern systems process and filter the data from the sensors to derive a coherent, recursive estimate of the AUV pose. This section will review some of the most common underwater sensors, popular state estimation filters, the basics of SLAM, and the foundations of cooperative navigation. A. Commonly Used Underwater Navigation Sensors Tables I-V describe some commonly used sensors for underwater navigation. Description Performance Cost A compass provides a globally bounded heading reference. A typical magnetic Accuracy On the order compass does so by measuring the magnetic field vector. This type of compass within 1 o of hundreds is subject to bias in the presence of objects with a strong magnetic signature and to 2 o for of dollars points to the earth s magnetic north pole. More common in marine applications, a modestly US. a gyrocompass measures heading using a fast spinning disc and the rotation priced unit. of the earth. It is unaffected by metallic objects and points to true north. TABLE I COMPASS Description Performance Cost Since the pressure gradient is much steeper underwater (10m = 1 atmosphere) we can achieve high accuracy 0.1m. Underwater depth can be measured with a barometer or pressure sensor. TABLE II PRESSURE SENSOR $ U SD Description Performance Cost Nominal standard deviation on The DVL uses acoustic measurements to capture bottom tracking and determine the velocity vector of an AUV moving across the seabed. It determines the AUV surge, sway, and heave velocities by transmitting acoustic pulses and measuring the Doppler shifted returns from these pulses off the seabed. DVLs will typically consist of 4 or more beams. 3 beams are needed to obtain a 3D velocity vector. TABLE III DOPPLER VELOCITY LOG the order of 0.3cm/s-0.8cm/s. $20k - 80k USD

4 4 Description Performance Cost A sonar is a device for remotely detecting and Along-track image resolution for an imaging Prices vary locating objects in water using sound. Passive side-scan sonar is a function of many factors widely from sonars are listening devices that record the such as range, sonar frequency, and water conditions, $20k sounds emitted by objects in water. Active sonars are devices that produce sound waves of specific, controlled frequencies, and listen for the echoes of these emitted sounds returned from remote objects in the water. Active sonars can be categorized as either imaging sonars that produce an image of the seabed, or ranging sonars which produce bathymetric maps. More details of specific active sonar devices are presented in Table IV in Sec. V-B. however cross-track resolution is inde- pendent of range. For example, a Klein 5000 side-scan operating at 455kHz can achieve an along track resolution of 10cm at 38m range and 61cm at the maximum 250m range and a Klein 5900 sidescan operating at 600kHz can achieve along track resolution 5cm at 10m and 20cm resolution at the maximum 100m range. In both cases nominal cross-track resolution is 3.75cm Resolution for a bathymetry sonar is on the order of 0.4 o 2 o along track and 5 10cm cross track [4]. 200kUSD or more. [4] TABLE IV SONAR Description Performance Cost Global Positioning Many factors can influence the accuracy of a GPS reading, including From Systems can be used type of GPS technique used, atmospheric conditions, number of hundreds for surface vehicles. satellites in view, and others. Precisions for different GPS systems are: to thousands Position is estimated common commercial off-the-shelf GPS - 10m, Wide Area Differential of dollars. using the time-offlight GPS (WADGPS) m, Real-Time Kinematic (RTK) m, of signals from Post processed m. synchronized satellites. TABLE V GLOBAL POSITIONING SYSTEM B. State Estimation The basis of any navigation algorithm is state estimation. Consider a robot whose pose at time t is given by x t. The goal of recursive state estimation is to estimate the belief distribution of the state x t denoted by bel(x t ) given by: bel(x t ) = p(x t u 1:t,z 1:t ) (1) where u is some control input or odometry and z is a measurement used for localization. The propagation of the state is given by some general non-linear process equation: x t = f(x t 1,u t,ǫ t ) (2) where ǫ t is process noise. The state is observable through some measurement function: z t = h(x t,δ t ), (3) where δ t is measurement noise. Typically, the state at time t is recursively estimated through an approximation of the Bayes filter which operates in a predict-update cycle. Prediction is given by [6]: bel(x t ) = x t 1p(x t x t 1,u t )bel(x t 1 ) (4)

5 5 Description Performance Cost Gyroscope Extremely - Drift variable. extremely From variable from hundreds o /hr of dollars Use a combination of accelerometers and gyroscopes (and sometimes magnetometers) to estimate a vehicle s orientation, velocity, and gravitational forces. Gyroscope: Measures angular rates. For underwater applications, the following two categories are widely used: Ring Laser / Fibre Optic: Light is passed either through a series of mirrors (ring laser) or fibre optic cable in different directions. The angular rates are determined based on the phase change of the light after passing through the mirrors or fibre. MEMS: An oscillating mass is suspended within a spring system. Rotation of the gyroscope results in a perpendicular Coriolis force on the mass which can be used to calculate the angular rate of sensor. Since a gyroscope measures angular rates, there will be a drift in the estimated Euler angles as a result of integration. Accelerometer: Measures the force required to accelerate a proof mass. Common designs include pendulum, MEMS, and vibrating beam among others. TABLE VI INERTIAL MEASUREMENT UNIT (RLG) to 60 o /hr or more for MEMS [5]. Accelerometer - Bias range from 0.01mg (MEMS) to 0.001mg (Pendulum) [5]. for MEMS IMU hundreds of thousands a to of dollars for a commercial grade ringlaser or fibre optic system. and update is given by: bel(x t ) = ηp(z t x t ) bel(x t ). (5) where η is a normalization factor. Implicit in this formulation is the Markov assumption, which states that only the most recent state estimates, control and measurements need to be considered to generate the estimate of the next state. Some of the more popular state estimation algorithms are summarized in Table VII. For further details, see for example [6]. All of the filters described in Table VII have been used in AUV navigation algorithms that will be described in the following sections. Some implementations differ by what variables are maintained in the state space as relevant to the navigation problem. For example, tide level [8], water current [9], [10], the speed of sound in water [11], or inertial sensor drift [11] can all be estimated to improve navigation. There are also popular variants of these classical filters. For example, since acoustic propagations are relatively slow compared to radio frequency communications it is often necessary to implement a delayed-state filter to account for the delay. Examples include [12] where a delayed state EIF is used and [13] which implements a delayed-state EKF. Often state estimation is decomposed into two parts: attitude heading and reference system (AHRS) and inertial navigation system (INS) as shown in Fig. 2. All sensors that give information about Euler angles or rates are inputs into the AHRS, which produces a stable estimate of vehicle orientation. The stabilized roll, pitch, and yaw are then used by the INS in combination with other sensors that give information about vehicle position, linear velocity, or linear acceleration to estimate the vehicle position. C. Simultaneous Localization and Mapping Simultaneous localization and mapping (SLAM) is the process of a robot autonomously building a map of its environment and, at the same time, localizing itself within that environment 1. SLAM algorithms can be either online, where only the current pose is estimated along with the map, or full where the posterior is calculated over the entire robot trajectory [6]. Analytically, online SLAM involves estimating the posterior over the momentary pose, x t, and the map, m given all measurements, z 1:t, and inputs u 1:t 1 also referred to as concurrent mapping and localization

6 6 State Estimator Bayes Filter Kalman Filter (KF) Extended Kalman Filter (EKF) Unscented Kalman Filter (UKF) Extended Information Filter (EIF) Particle Filter (PF) Least Squares Regression Description Optimal but computationally intractable for all but the simplest of estimation problems. State distribution assumed to be Gaussian and parameterized by mean, µ, and covariance, Σ. Requires (2) and (3) to be linear. Optimal conditional on Gaussian and linearity assumptions. Extension of the Kalman filter to accommodate non-linear process and measurement models. On each iteration, (2) and (3) are linearized about the mean. Prediction operation is fast, but measurement update is slow as it requires matrix inversion. Reduces the linearization errors of the EKF at the expense of higher computation. Instead of just mean and covariance being mapped through the non-linear functions, multiple sigma points are mapped and then the output is re-parameterized as Gaussian. State distribution assumed Gaussian but parameterized by information matrix, I = Σ 1 and the information vector, ζ = Σ 1 µ. Sometimes referred to as the canonical representation. Allows for processing of multiple measurements at one time easily though addition. Prediction step can be slow as it requires matrix inversion, but measurement update step is fast. Also can be time consuming to recover the mean and covariance. However, in some cases has certain advantages over the EKF. Non-parametric representation of state distribution. Instead distribution is represented by discrete particles with associated weights. Has advantage that non- Gaussian distributions and non-linear models can be incorporated. Computation scales with the number of particles in the particle set. Obtaining the Maximum A Posteriori state estimate can be formulated as a least squares optimization [7] which can be solved analytically. This formulation has the advantage that past states are maintained which can be useful for full trajectory optimization or simultaneous localization and mapping. TABLE VII SOME COMMON STATE ESTIMATION TECHNIQUES IMU, compass Attitude Heading and Reference System stabilized roll, pitch, yaw Inertial Navigation System position GPS, LBL, USBL, SBL, SLAM, DVL, acoustic range, depth sensor, ADCP, others Fig. 2. Position estimation with an attitude heading and reference system and an inertial navigation system. p(x t,m z 1:t,u 1:t ), (6) whereas full SLAM involves estimating the posterior over the entire pose trajectory p(x 1:t,m z 1:t,u 1:t ). (7)

7 7 a) b) L 1 L 8 L 9 V 1 V 2 V 3 V 4 L 4 L 2 P 1 P 2 P 3 P 4 L 6 P 1 P 2 P 3 P 4 L 3 L 5 L 7 Fig. 3. (a) Feature based SLAM, (b) View based SLAM. In addition, SLAM implementations can be classified as feature based, where features are extracted (detection, identification and classification) and maintained in the state space, or view based, where poses corresponding to measurements are maintained in the state space. As Fig. 3-a shows, in feature based SLAM features are extracted from sensor measurements. For example, at pose P 1 the robot sees three features L 1, L 2 and L 3. These features together with the pose of the robot are maintained in the state space. At the next pose, P 2, only newly observed features, L 4 and L 5 are added to the vector and the pose is replaced with the previous pose. This process occurs at each new pose. In view based SLAM (Fig. 3-b), at each pose the whole view without extracting any features is processed usually by comparing it with the previous view. For example, at pose P 3, V 3 is compared with V 2 to find the view based odometry. State vectors in this case can be composed by one or more of the poses at each time. SLAM method Pros Cons AUV Application EKF SLAM [14] Works well when features are present and distinct (which can be challenging underwater ). SEIF SLAM [23] Performs updates in constant time. Due to additivity of information, it is a good choice for multiple-robot SLAM. FastSLAM [27] Logarithmic time in number of features. No dependance on parametrization of motion models. GraphSLAM [33] Previous poses are updated for post-processing of data. AI SLAM Efficient, because it mimics the way animals brain work. Adding new features to state space requires quadratic time. Information matrix has to be actively sparsified. Recovering map requires matrix inversion. Ability to close loops depends on particle set. More computation required. Covariances are hard to recover (information form). Requires training or parameter tuning. TABLE VIII STRENGTHS AND WEAKNESSES OF SOME COMMON SLAM TECHNIQUES [15] [16] [17] [18] [19] [20] [21] [22] [24] [25] [26] [28] [29] [30] [31] [32] [34] [35] [36] Filtering (online) approaches to SLAM make use of a state estimation algorithm such as those presented in Table VII. Smoothing (full SLAM) methods, also known as GraphSLAM [33], minimize the process and observation constraints over the whole trajectory of the robot. Some approaches use a combination of methods. Some of the most popular categories of SLAM approaches are described here with their pros, cons, and AUV

8 8 navigation references provided in Table VIII. The categorization is based on [6] with some additions: EKF SLAM: EKF-SLAM linearizes the system model using the Taylor expansion. It applies recursive predictupdate cycle to estimate pose and map. Its state vector includes pose and features [14]. It is applicable to both view based SLAM [37] and feature based SLAM [38]. For large maps, EKF-SLAM is computationally expensive since computation time scales O(n 2 ) where n is the number of features. SEIF SLAM: Sparse extended information filter (SEIF) [23] and exactly sparse extended information filters (ESEIF) [39] are two well-known approaches for SLAM using the information filter. They both maintain a sparse information matrix which preserves the consistency of the Gaussian distribution; however, accessing the mean and covariance requires a computationally expensive large matrix inversion. Both approaches need the information matrix to be actively sparsified by a sparsification strategy. ESEIF maintains an information matrix with the majority of elements being exactly zero which avoids the overconfidence problem of [23]. FastSLAM: FastSLAM is based on the particle filter. Particle filtering approaches are nonlinear filtering solutions; therefore, the system models are not approximated. In FastSLAM, poses and features are represented by particles (points) in the state space [27]. FastSLAM is the only solution which performs online SLAM and full SLAM together, which means it estimates not only the current pose, but also the full trajectory. In FastSLAM, each particle holds an estimate of the pose and all features; however, each feature is represented and updated through a separate EKF. Similar to other methods, it is applicable to both view based SLAM [40] and feature based SLAM [6]. GraphSLAM: In GraphSLAM methods, the entire trajectory and map are estimated [33]. GraphSLAM also uses approximation by Taylor expansion, however it differs from EKF-SLAM in that it accumulates information and therefore is considered to be an off-line algorithm [6]. Generally, in GraphSLAM, poses of the robot are represented as nodes in a graph. The edges connecting nodes are modeled with motion and observation constraints. These constraints need to be optimized to calculate the spatial distribution of the nodes and their uncertainties [6]. Different solutions exist for GrpahSLAM such as relaxation on a mesh [41], multi-level relaxation [42], iterative alignment [43], square root smoothing and mapping (SAM) [7], incremental smoothing and mapping (isam) [44] and works by Grisetti et. al in [45], [46], and hierarchical optimization for pose graphs on manifolds (HOGMAN) [47]. In principle, they are all similar, but differ in how the optimization is implemented. For instance, isam solves the full SLAM problem by updating a matrix factorization while HOGMAN s optimization is performed over a manifold. Artificial Intelligence (AI) SLAM: These methods of SLAM are based on fuzzy logic and neural networks. ratslam [48] is a technique that models rodents brain using neural networks. In fact this method is neural network-based data fusion using a camera and an odometer. [49] uses self organizing maps (SOM) to perform SLAM with multiple-robots. The SOM is a neural network which is trained without supervision. The choice of the method for estimating the poses of robots and the map depends on many factors such as the available memory, processing capability and type of sensory information. SLAM techniques have been used for acoustic (Sec. IV) and particularly geophysical (Sec. V) underwater navigation algorithms as will be described. D. Cooperative Navigation In cooperative navigation (CN), AUV teams localize using proprioceptive sensors as well as communications updates from other team members. CN finds its origin in ground robotics applications. In the seminal paper by Roumeliotis and Bekey [50], it is proven that a group of autonomous agents with no access to global positioning can localize themselves more accurately if they can share pose estimates and uncertainty as well as make relative measurements. In [51], the scalability of CN is addressed, and it is shown that an upper bound on the rate of increase of position uncertainty is a function of the size of the robot team. Other important results have been proven, such as that the maximum expected rate of uncertainty increase is independent of the accuracy and number of inter-vehicle measurements and depends only on the accuracy of the proprioceptive sensors on the robots [52]. In addition, applications of Maximum A Posteriori [53], [54], EKF [55] and nonlinear least squares [56] estimators have been developed for general robotics CN. A complexity analysis is also presented in [57]. Special considerations must be made to apply

9 9 σ y2,y2 σ y1,y1 R 12,R 21 σ x2,x2 x2, y2, z2 x1, y1, z1 σ x1,x1 R 13,R 31 R 23,R 32 Actual Position Estimated Position σ x3,x3 x3, y3, z3 σ y3,y3 Fig. 4. Cooperative navigation for AUVs: relative ranges are determined from TOF of acoustic communication packets. many of these algorithms to underwater CN since the acoustic communications channel is limited. Further detail will be presented in Sec. IV. A graphical depiction of multiple AUV CN is shown in Fig. 4. Data is transmitted through the acoustic channel. Upon reception of a data packet, the receiver, vehicle j, can use the time-of-flight of the acoustic signal to determine its range, R ij, from the sender, vehicle i. If the vehicles possess well synchronized clocks, then this range can be determined from the one-way travel time (OWTT) of the acoustic signal, otherwise an interrogation-reply is performed to determine a round-trip range (RTR). Ranges are usually projected onto the xy plane to obtain R xy ij since the depths of both vehicles, z i and z j, are observable with pressure sensors: The range measurement model, h is given by: R xy ij = (R2 ij (z i z j ) 2 ) 1/2. (8) h ij (X i,x j ) = ((x i x j ) 2 +(y i y j ) 2 ) 1/2 +v R (9) where X i and X j are the states of vehicles i and j with positions (x i,y i ) and (x j,y j ) respectively. The noise term v R N(0,σR 2 ) is zero-mean Gaussian noise. The measurement model is nonlinear and assumes that error in the range measurement, v R, is independent of range. Some previous work has validated this assumption to some extent [58]. In [35], it is assumed that σ R = 3m is a reasonable value if using OWTT, and σ R = 7m is reasonable for RTR. An important consideration for AUV CN, as with any CN algorithm, is to track the cross-correlations that are induced from inter-vehicle measurements in order to avoid overconfidence in the estimate. III. INERTIAL When the AUV positions itself autonomously, with no acoustic positioning support from a ship or acoustic transponders, it dead reckons. With dead reckoning (DR), the AUV advances its position based upon knowledge of its orientation and velocity or acceleration vector. Traditional DR is not considered a primary means of navigation but modern navigation systems, which depend upon DR, are widely used in AUVs. The disadvantage of DR is that errors are cumulative. Consequently, the error in the AUV position grows unbounded with distance traveled.

10 10 One simple method of DR pose estimation, for example if heading is available from a compass and velocity is available from a DVL, is achieved by using the following kinematic equations: ẋ = vcosψ +wsinψ ẏ = vsinψ +wcosψ ψ = 0 where (x, y, ψ) is the displacement and heading in the standard North-East-Down coordinate system, and v, and w are the body frame forward and starboard velocities. In this model it is assumed that roll and pitch are zero and that depth is measured accurately with a depth sensor. An inertial system aims to improve upon the DR pose estimation by integrating measurements from accelerometers and gyroscopes. Inertial proprioceptive sensors are able to provide measurements at a much higher frequency than acoustic sensors that are based on the TOF of acoustic signals. As a result, these sensors can reduce the growth rate of pose estimation error, although it will still grow without bound. One problem with inertial sensors is that they drift over time. One common approach, for example used in [11], is to maintain the drift as part of the state space. Slower rate sensors are then effectively used to calibrate the inertial sensors. In [11], the authors also track other possible sources of error such as the variable speed of sound in water to reduce systematic noise. These noise sources are propagated using a random walk model, and then updated from DVL or LBL sensor inputs. Their INS is implemented with an IMU that runs at 150Hz. The basic kinematics model (10) is incomplete if the local water current is not accounted for. The current can be measured with an acoustic Doppler current profiler (ADCP). For implementations with ADCP see [9] [10]. A DVL is usually able to calculate the velocity of the water relative to the AUV, v b and the velocity of the seabed relative to the AUV, v g. Then the ocean current can be calculated easily as v c = v g - v b. The ocean current can also be obtained from an ocean model, for example in [59] where ocean currents are predicted using the regional ocean modeling system [60] combined with a Gaussian process regression [61]. If access to the velocity over the seabed is not available, then the current can be estimated from a transponder on a surface buoy as in [62]. In [62], the authors analyzed the power spectral density to remove the low frequency excitation on the buoy due to the waves to estimate the underwater current. In [8], an algorithm based on particle filtering is proposed that exploits known bathymetric maps - if they exist. It is emphasized that the tide level must be carefully monitored to avoid position errors, particularly in areas of low bathymetric variation. This approach is referred to as terrain-aided navigation, and the method is compared for DVL, and multi-beam sonars as the bathymetric data input, concluding that both are viable options. The performance of an INS is largely determined by the quality of its inertial measurement units. In general, the more expensive the unit, the better its performance. However, the type of state estimation also has an effect. The most common filtering scheme is the EKF, but others have been used to account for the linearization and Gaussian assumption shortcomings of the EKF. For example, in [63] a UKF is used and in [64] a PF application is presented. Improvements can also be made to INS navigation by modifying Eq. (10) to provide a more accurate model of the vehicle dynamics. The benefits of such an approach are investigated in [65], particularly in the case that DVL loses bottom lock, for example. Inertial sensors are the basis of an accurate navigation scheme, and have been combined with other techniques described in subsequent sections. In certain applications, navigation by inertial sensors is the only option. For example, in extreme depths where it is impractical to surface for GPS, an INS is used predominantly, as described in [66]. The best INS can achieve a drift of 0.1% of the distance traveled [35], however, more typical and modestly priced units can easily achieve a drift of 2-5% of the distance traveled. (10) IV. ACOUSTIC TRANSPONDERS AND BEACONS In acoustic navigation techniques, localization is achieved by measuring ranges from the TOF of acoustic signals. Common methods include: Ultra Short Baseline (USBL): Also sometimes called super short baseline (SSBL). The transducers on the transceiver are closely spaced with the approximated baseline on the order of less than 10 centimeters. Relative

11 11 Fig. 5. (a) Short Baseline (SBL) (b) Ultra-Short Baseline (USBL) (c) Long Baseline (LBL) ranges are calculated based on the TOF and the bearing is calculated based on the difference of the phase of the signal arriving at the transceivers. See Fig. 5-b. Short Baseline (SBL): Beacons are placed at opposite ends of a ship s hull. The baseline is based on the size of the support ship. See Fig. 5-a. Long Baseline (LBL) and GPS intelligent Buoys (GIBs): Beacons placed over a wide mission area. Localization is based on triangulation of acoustic signals. See Fig. 5-c. In the case of GIBs, the beacons are at the surface whereas for LBL they are installed on the seabed. Single Fixed Beacon: Localization is performed from only one fixed beacon. Acoustic Modem: The recent advances with acoustic modems have allowed for new techniques to be developed. Beacons no longer have to be stationary, and full AUV autonomy can be achieved with support from autonomous surface vehicles, equipped with acoustic modems, or by communicating and ranging in underwater teams. Due to the latency of acoustic updates, state estimators are implemented where the dead reckoning proprioceptive sensors provide the predictions and then acoustic measurements provide the updates. A. Ultra Short and Short Baseline Ultra short baseline (USBL) navigation allows an AUV to localize itself relative to a surface ship. Relative range and bearing are determined by TOF and phase differencing across an array of transceivers, respectively. A typical setup would be to have a ship supporting an AUV. In short baseline (SBL), transceivers are placed at either end of the ship hull and triangulation is used. The major limitation of USBL is the range and of SBL is that the positional accuracy is dependent on the size of the baseline, i.e. the length of the ship. In [67] an AUV was developed to accurately map and inspect a hydro dam. A buoy equipped with an USBL and differential GPS helps to improve upon dead reckoning of the AUV which is performed using a motion reference unit (MRU), a fibre optic gyro (FOG) and a DVL. An EKF is used to fuse the data and a mechanical scanning imaging sonar (MSIS) tracks the dam wall and follows it using another EKF. For this application, the USBL is a good choice because the range required for the mission is small. The method proposed in [12] augments [67] by using a delayed-state information filter to account for the time delay in the transmission of the surface ship position. In [68], sensor based integrated guidance and control is proposed using a USBL positioning system. The USBL is installed on the nose of the AUV while there is an acoustic transponder installed on a known and fixed position as a target. While homing, the USBL sensor listens for the transponder and calculates its range and the bearing based on the time difference of arrival (TDOA). In [69], USBL is used for homing during the recovery of an AUV through sea ice. In [70], two methods are presented to calibrate inertial and DVL sensors. The inertial navigation system data from the AUV is sent to the surface vehicle by acoustic means. In one method a simple KF implementation is used which maintains the inertial sensor drift errors in the state space. In the other method, possible errors of the USBL in the sound velocity profile are incorporated and the EKF is used to fuse data. No real hardware implementation is performed. In [71], the method is extended to multiple AUVs by using an inverted setup where the the transceiver is mounted on the AUV and the transponder mounted on the surface ship.

12 12 In [72], data from an USBL and an acoustic modem is fused by a particle filter to improve dead reckoning. As a result, the vehicle operates submerged longer as GPS fixes can be less frequent. The simulation and field experiments verify the developed technique. In [73], a tightly-coupled approach is used where the spatial information of the acoustic array is exploited to correct the errors in the INS. B. Long Baseline / GPS Intelligent Buoys In LBL navigation, localization is achieved by triangulating acoustically determined ranges from widely spaced fixed beacons. In most cases, the beacons are globally referenced before the start of the mission by a surface ship [74], a helicopter [75], or even another AUV [76]. In normal operation, an AUV would send out an interrogation signal, and the beacons would reply in a predefined sequence. The two way travel time (TWTT) of the acoustic signals is used to determine the ranges. However, there have been implementations in which synchronized clocks are used to support one-way travel time (OWTT) ranging [77]. GIBs remove the need for the LBL beacons to be installed at the seafloor which can reduce installation costs and the need for recovery of these beacons. One of the limitations of LBL is the cost and time associated with setting up the network. However, this can be mitigated to some extent if the beacon locations are not globally referenced and either self-localize [78], or the AUV can localize them by performing SLAM. For example, [34] uses a nonlinear least squares implementation, whereas [79] uses a particle filter version of SLAM to determine the location of the fixed beacons during the mission. A major consideration in an LBL localization network is the treatment of outliers. Methods to account for outliers in LBL systems include hypothesis grids [80] and graph partitioning [81]. Generally, range measurements can fall into one of three categories: direct path (DP), multi-path (MP), or outlier (OL). From experience, range errors are not Gaussian distributions. The quality of range data is dependent on the location within the survey area. In [80], a hypothesis grid is built to represent the belief that future measurements from a particular cell will be in a particular category (i.e. DP, MP, OL). In graph partitioning, outliers are rejected using spectral analysis. A set of measurements is represented as a graph, then the graph partitioning algorithm is applied to identify sets of consistent measurement [81]. Another consideration is the TDOA of the acoustic responses of the network [82], [83]. The change in vehicle pose between the initial interrogation request and all of the subsequent replies must be explicitly handled. This is often done with a delayed state EKF. Each range difference measurement between two receivers constrains the target to a annulus (in 2D) or a sphere (in 3D). Annuluses are intersected to find the location of the target. However, when the data is corrupted by noise, there is not necessarily an intersection point. Sequential quadratic programming [84] is used to perform the constrained optimization. The target position estimate is the same as the traditional maximum likelihood estimate. Major drawbacks of LBL are the finite range imposed by the range of the beacons, and the reliance on precise knowledge of the local sound velocity profile of the water column based on temperature, salinity, conductivity, and other factors [74]. However, the LBL systems do overcome these shortcomings to be one of the most robust, reliable, and accurate localization techniques available. For that reason it is often used in high-risk situations such as under-ice surveys [75], [85]. Other implementations include: alignment of Doppler sensors [86] and deep water surveys [87]. An extensive evaluation of the precision of LBL is provided in [88]. As a surveying implementation example, in [87] a technique is used for deep sea near bottom survey by an AUV which relies on LBL for navigation. LBL transponders are GPS-referenced before the start of the survey. In [86], two techniques for in situ 3 degrees-of-freedom calibration of attitude and Doppler sonar sensors are proposed. LBL and gyrocompass measurements are sources of information for alignment. C. Single Fixed Beacon A downside of LBL systems is the cost and time required for installing the beacons and geo-referencing them. It is possible to reduce these infrastructure requirements if only a single fixed beacon is used instead of a network of them. The concept is that the baseline is simulated by propagating the ranges from a single beacon forward in time until the next update is received. This technique has been referred to as virtual LBL (VLBL) and has been simulated

13 13 Fig. 6. An AUV localizing with a single fixed beacon at known location. Uncertainty grows in between updates from the beacon. On reception of an update from the beacon, uncertainty is reduced in the dimension coinciding with the location of the beacon. on real world data in [89]. It is noted that the AUV trajectory has a significant effect on the observability of the vehicle state. Long tracks directly towards or away from the single fixed beacon will cause unbounded growth in position error. As a result, tracks or paths for the survey vehicle being localized in this manner should be planned to be tangential to range circles emanating from the transponder. A visual representation of single beacon navigation is shown in Fig. 6. It assumes that the vehicle has prior knowledge of the beacon location. In the figure the vehicle receives 3 acoustic pings from the beacon at the bottom. Each time, reception of a ping results in a reduction of uncertainty in the direction of the beacon. An important consideration when the baseline is removed is observability of the state space. As detailed in [90], and elaborated upon in [91] and [92], if the states are estimated by a linearized filter, such as an EKF or EIF, then observability is lost if repeated range updates come from the same relative bearing. However, if the states are estimated with a nonlinear filter, then the condition for loss of observability changes to repeated range updates coming from the same relative bearing and the same range. As a result of the nonlinear nature of the range update model, it is beneficial to use a nonlinear state estimator from an observability standpoint. This is derived in detail in [91] and [92] using Lie derivatives to generate the observability matrix. Single beacon navigation has also been used for homing, for example in [93]. This task is particularly challenging because the AUV will preferentially move in a track directly towards the beacon, violating the observability criterion. As a result, path planning must be designed to avoid this situation. This can be particularly useful for recovering an AUV that has become inoperable, but is still able to transmit pings. D. Acoustic Modem Advances in the field of acoustic communications have had a major effect on underwater navigation capabilities. The acoustic modem allows simultaneous communication of small packets and ranging based on TOF. If the position of the transmitter is included in the communicated information, then the receiver can bound its position to a sphere centered on the transmitter. This capability removes the need for beacons to be fixed or localized prior to the mission. In addition, it allows for inter-auv communication, which means teams of AUVs can cooperatively localize. Popular acoustic modem are manufactured by Woods Hole Oceanographic Institute [94], Teledyne Benthos [95], and Evologics [96] among others. In general communication can either use frequency shift keying with frequency hopping (FH-FSK), which is more reliable but provides lower data rates, or variable rate phase-coherent keying (PSK). Some models also include precise pulse-per-second (PPS) clocks (eg [97]) to allow synchronous ranging. Typically, due to the limited bandwidth underwater, the communication channel is shared using a time-division multiple access (TDMA) scheme. In TDMA, each member in the group is allotted a time slot with which to broadcast information. The major detractor of such a scheme is that the total cycle time grows with group size. At present, achievable bit rates range from 32 bytes per 10s packet with FSK, to several kbit/s in optimal conditions with PSK. 1) Manned Surface Support: The ability of a modem at the surface to transmit its location to the survey vehicles provides two important benefits over past navigation methods: 1) it removes the necessity to geo-reference the

14 beacons prior to starting the mission, and 2) It allows the beacons to move during the missions. The first advantage saves time and money, and the second allows the mission range to be extended as necessary without re-deploying the sensor network. Many methods have been recently published that exploit one or both of these benefits. The moving long baseline (MLBL) concept was first demonstrated in [98] using a Sonardyne AvTrack acoustic navigation system. Two manned surface vehicles were used to support one AUV. This concept has proved particularly useful for mapping rivers, such as [99]. In this project, two boats are used to continuously define a cross-section of the river to be mapped near the location of the AUV. This approach has been extended to a single moving source. A manned surface ship can localize itself with GPS and bound the error of one or more survey AUVs by broadcasting its position. Such an approach is attractive because there is no need for calibration or recovery of beacons. The approach is similar to that presented in Sec. IV-C except the mission range can be much larger because the surface vehicle can move. In [100] and [13], a deep water validation was performed for the single moving beacon concept. It should be noted that localization is done in post-processing. In the proposed approach, an EKF maintains an estimate of the survey vehicle as well as the support ship. In a deep water application, the time taken for the acoustic transmission should be accounted for in the filtering algorithm. Here, this is represented as a delayed state EKF. It is accurately noted that the TOF measured is the range between the current position of the receiver and a previous position of the sender. In [13], the performance of the single beacon navigation is compared against an LBL system. Also, there is a more rigorous discussion on sources of error in acoustic range measurements such as errors in sound velocity estimation, acoustic multi-path, and errors in ship GPS. Similarly, in [101] a deep diving AUV is localized in post-processing. In this case, the survey vehicle is required to execute a known closed path after it dives to a depth of 6000m. A similar approach is presented in [102] to map the magnetic signature of a moving vessel. If the survey vehicles are acting as passive listeners using OWTT for range measurement, then the system naturally scales well with number of survey vehicles assuming they stay within range of the surface vehicle. In [103] and [104], a maximum likelihood sensor fusion technique is presented to localize survey vehicles to within 1m over a 100km survey. It is also noted in this paper that over a long survey, the drift of the PPS clock will have a significant effect on the localization performance. 2) Autonomous Surface Crafts: Once it is established that AUVs can navigate with the help of manned surface vehicles, a natural progression to increase autonomy is to move towards unmanned surface vehicles. The first known implementation of autonomous surface crafts (ASCs) used to support AUVs is presented in [58], which is an extension of the MLBL concept presented in [98]. Two ASCs are used to support one Odyssey III AUV in a series of experiments in 2004 and In [105], two ASCs are used and a general framework is developed for cooperative navigation. Following previous trends, the logical progression was to perform AUV navigation and localization with only one ASC. This has been experimentally shown in [91] and [106]. The approaches taken were similar, but in [106], no actual AUVs were used in the implementation, instead an ASC was used as a surrogate. In both cases, the authors implemented a nonlinear least squares (NLS) approach. In [91], the experimental validation is done using an ASC and an AUV to compare the performance of the EKF, PF, and an NLS optimization. It is shown that the NLS performs the best, particularly after off-line post-processing. The observability is carefully considered in both cases. In [91], it is shown that the observability criterion is less stringent if the states are being estimated with a nonlinear filter. In addition, the ASCs autonomy includes basic heuristics for supporting the survey AUV and maintaining observability by performing different motion behaviors. In addition, in [91], an error analysis of the range updates is performed, and the noise found to be relatively close to normally distributed and independent of range. Path planning of the ASC to maintain observability is also considered explicitly in [107]. More recently, an observability analysis was done for an ASC/AUV team under the cases that position, and that position and velocity, are transmitted from the ASC to the AUV [108]. A method for globally asymptotically stable observer design is also presented. In [109], an AUV is able to localize relative to a surface ship or a drifter whose position is estimated with the use of an ocean current model. Furthermore, the AUV is able to obtain relative range and bearing estimates by detecting the surface ship or buoy using an upward-looking sonar. A theoretical analysis of the benefits of such an approach as well as performance bounds are derived. It is expected that in the coming years, ASCs, such as the wave glider [110], will be capable of providing extremely long-term autonomous navigation capabilities. One such example scenario is presented in [111] where a 14

15 15 wave glider is used to localize an active acoustic source. 3) Heterogeneous AUV Teams: In some cases, certain AUVs are outfitted with more expensive sensors and/or make frequent trips to surface for GPS position fixes. These vehicles support the other survey vehicles and have been referred to as communications/navigations aids (CNAs) 2. An application would be to have inferior navigation sensors on a fleet of survey vehicles and expensive navigation sensors on only a small number of vehicle whose job is to support that fleet [112]. Upon reception of a CNA transmission the receiver s position estimate can be restricted to an annulus if each vehicle has a good estimate of depth (which can easily be obtained from a pressure sensor), and if the receiver can estimate its range to the CNA, for example through either one-way TOF of the acoustic signal (requires synchronized clocks) or two-way TOF. This annulus can be intersected with a vehicle s own covariance ellipse as shown in Fig. 7. ` Fig. 7. A CNA AUV supporting a survey AUV [112]. The blue areas are the vehicle uncertainty ellipses. The yellow area is the updated uncertainty of the survey ellipse from a range measurement, which is shown in pink. The framework for cooperative navigation developed by Bahr and others is presented in [113], [105], [114], [115] and is applicable to general AUV teams. Referring to Fig. 7, each annulus is propagated forward in time until the next update is received from the CNA. In between updates, the uncertainty grows based on the quality of the DR estimation. The importance of careful book-keeping and maintenance of cross-correlations between survey and CNA AUVs is stressed to avoid over-confidence and to enhance the quality of outlier rejection. While the developed framework is to be used for pure AUV systems, all results show the use of ASCs as surrogate AUVs. An altogether different approach to CN is presented in [116] where a hovering type AUV can physically dock to the torpedo type AUV in order to exchange information. 4) Homogeneous AUV Teams: When a homogeneous team performs CN, each member of the team is treated equally. Position error will grow slower if the AUVs are able to communicate their relative positions and ranges. Similarly to the heterogeneous case, if any vehicle surfaces for a position fix, then the information can be shared with the rest of the team to bound the position error. The first known experimental implementation of such a system was in 2007 [117]. In this approach, the vehicle states are maintained with an EKF but cross covariances are not transmitted in the communications packet due to bandwidth limitations. The communication latency is accounted for in the EKF implementation. More recently, an implementation of CN was presented in [118]. Here, the issue of book-keeping is explicitly addressed. A multi-vehicle ledger system is used so updates are only made to the multi-vehicle EKF when all up to date data has been obtained. Experiments were performed with kayaks acting as AUV surrogates, where OWTT ranging was done with the synchronized clocks in a 30s cycle, each of the three vehicles having a 10s slot to make a transmission. V. GEOPHYSICAL Geophysical navigation refers to any method that utilizes external environmental features for localization. Almost all methods in this category that achieve bounded position error use some form of SLAM. Categories include: 2 This term has also been applied to surface vehicles supporting AUV survey teams

16 16 Magnetic: It has been proposed to use magnetic field maps for localization. Although no recent publications have been found, a team at the University of Idaho has been mapping the magnetic signatures of Navy vessels [119], [120]. Optical: Use of a monocular or stereo camera to capture images of the seabed and then match these images to navigate. Sonar: Used to acoustically detect then identify and classify features in the environment that could be used as navigation landmarks. With bathymetric sonar features can be extracted almost directly from assembled returns. With sidescan (imaging) sonar, feature extraction is achieved through processing of imagery. A. Optical Visual odometry is the process of determining the robot pose by analyzing subsequent camera images. This can be achieved through optical flow or structure from motion (SFM). Invariant extraction and representation of features is an important consideration. Many previous algorithms have been proposed and applied in ground and air robotics, such as scale-invariant feature transform (SIFT) [121], speeded up robust feature (SURF) [122], amongst many others 3. Images can be captured with either stereo or monocular cameras. Stereo cameras have the added advantage that full 6 degree-of-freedom transformations between consecutive image pairs can be found. Visionbased SLAM can also be performed with the methods presented in Table VIII. The major challenge is closing loops in the trajectory by associating non-consecutive images, which is necessary to bound localization error. Limitations for optical systems in underwater environments include the reduced range of cameras, susceptibility to scattering, and inadequacy of lighting. As a result, visible wavelength cameras are more commonly installed on hovering AUVs because they can get close to objects of interest. In addition, visual odometry and feature extraction relies on the existence of features. Therefore optical underwater navigation methods are particularly well-suited to small-scale mapping of feature-rich environments. Examples include ship hulls or shipwreck inspections. In [16] and [17], Eustice et al. present an implementation of underwater vision-based SLAM called visually augmented navigation (VAN). The multi-sensor approach combines the benefits of optical and inertial navigation methods and is robust to low overlap of imagery. The approach is a view based version of EKF-SLAM, where camera-derived relative pose measurements provide the spatial constraints for visual odometry and loop closure. In [24] and [25], the VAN approach is converted to the information form and it is proven that the view based ESEIF SLAM maintains a sparse information matrix without approximations or pruning. This approach was applied to deep water surveying of the RMS Titanic in [24]. The issue of recovering the mean and the covariance from the information form is also addressed. SIFT and Harris extraction points are used to match images. The VAN method has also been applied to ship-hull inspection for the US Navy [124]. Both [44] and [125] were inspired by the VAN method. However, these two methods improved upon the approach using a smoothing and mapping problem formulation and doing efficient matrix factorizations to be able to efficiently recover the means and covariances. In recent work, [126] uses features from both a profiling sonar and a monocular camera in full pose graph formulation for a hovering AUV performing ship hull inspections. Odometry is derived relative to the ship hull using a DVL locked onto the ship. An approach to tackle the problem of data association in feature-poor areas of a ship hull is explored in [127] using a novel online bag-of-words approach to determine inter-image and intra-image saliency. Feature based approaches to underwater SLAM have also been done. For example, in [18], an augmented EKF is used to generate a topological representation. Non-time-consecutive images are compared and loop closures are made based on observation mutual information. Feature based EKF SLAM is also applied to the underwater environment in [19] and [20]. A byproduct of accurate localization is that accurate mapping can be achieved. Several authors have shown the ability to compute 2D and 3D reconstructions of the underwater environment based on optical underwater SLAM. For example, in [128], an SFM model together with SLAM are used to photomosaic the RMS Titanic as well as a hydrothermal vent area in the mid-atlantic. In [129], 3D reconstructions of an underwater environment are done using SLAM with a combination of multi-beam sonar and stereo camera. Future research will likely involve incorporating multiple vehicles. Recently, some publications involving simulations report on this [130], [131]. Overcoming communications constraints will be a major challenge. 3 For a review and comparison of some methods, see [123].

17 17 Pure visual odometry based methods that do not require SLAM can be used for applications such as pipeline tracking [132]. B. Sonar Sonar imaging of the ocean pre-dates AUVs by decades. As a result, it is a fairly robust technology. Several types of sonars are used for seabed and structure mapping. They can be categorized as imaging-type which provide an image of the seabed (see Table IV) and ranging-type which produce range or bathymetric data (see Table X). Sonar Description Pros Trade-Off Application Sidescan multiple beams that measure intensity of returns to create a 2D image of sea bed; beams are directed perpendicular to can work at relatively high speeds (10kt) to give high area coverage. resolution inversely proportional to range, e.g. 1.8 [133] [35] [134] [135] travel direction. MHz produces Forward Look (imaging type) Synthetic Aperture Mechanical Scanned Imaging Sonar similar in principle to a sidescan sonar only beams are directed forward. coherent processing of consecutive displaced returns to synthesize a virtual array. One beam with actuator that scans a swath. obstacle avoidance, also as a nadir gap filler. range independent resolution Cheaper than multibeam. 40 m range. limited distance to depth ratio ( 6:1 max); single angle of view. optimal at low speeds and deeper water. Slow. Accuracy depends on AUV attitude. TABLE IX IMAGING-TYPE SONAR DEVICES USED FOR UNDERWATER NAVIGATION [26] [136] [137] [138] [139] [15], [21], [140], [141], [142], [143], [22] Sonar Description Pros Trade-Off Application Echo point measurements in generally [144], Sounder one direction. [145] single, narrow beam used to determine depth below the transducer. Profiler low-frequency echo sounders that penetrate the seabed. Multi- Beam time-of-flight from returns to assemble bathymetric maps. yield representation of sea bed and targets between transducer to seabed. information on subsurface features. gathers echo sounding data more efficiently than a single beam. penetration depth inversely proportional to resolution. resolution inversely proportional to frequency; may not always capture the first ping. TABLE X RANGING-TYPE SONAR DEVICES USED FOR UNDERWATER NAVIGATION [146] [28] [29] [30] [31] Sonars are designed to operate at specific frequencies [147] depending on the range and resolution required. In all cases, the performance of the SLAM algorithm is dependent on the number and quality of the features present in the environment. 1) Imaging Sonar: The insonified swath of the sidescan sonar is shown in Fig. 8-a. The intensity of the acoustic returns from the seabed of this fan-shaped beam also depends on the bottom type and is recorded in a series of cross-track slices. When mosaicked together along the direction of travel, these assembled slices form an image

18 18 a) b) c) d) e) Fig. 8. Sonar sensor swaths: (a) Sidescan (b) Multibeam (c) Foward looking (d) Mechanical scanning and imaging (e) Synthetic aperture. of the seabed within the swath of the beam. Hard objects protruding from the seabed send a strong return which is represented as a dark image. Shadows and soft areas, such as mud and sand, send weaker returns which are represented as lighter images. SLAM with the sidescan was first presented in [133] using an augmented EKF and the Rauch-Tung-Striebel smoother. The value of smoothing in the pose estimation is emphasized since all previous poses should be updated when a loop closure event is detected. It is noted that automated feature detection (which is not trivial) and data association are necessary to achieve autonomy for sidescan sonar SLAM. In [35], the approach in [133] is improved using isam for smoothing and incorporates range updates from a CNA. This is the only known research that combines a sonar based SLAM method with acoustic modem ranging. In [134], sidescan sonar SLAM is posed as an interval constraint propagation problem. In [148], [135], the problem was approached using a submap joining algorithm called the selective submap joining SLAM. The work in [135] uses a cascaded Haar classifier for object detection from sidescan imagery and is demonstrated off-line on a gathered dataset. A discussion of which type of features are appropriate for sidescan sonar SLAM is presented in [149]. All of the proposed approaches to sidescan sonar SLAM require post-processing to detect the features for data association. Presently, this is just becoming achievable on-board AUVs. A depiction of the forward looking sonar (FLS) is shown in Fig. 8-c. Based on the transducer geometry, the primary function of this type of sonar device is to map vertical features. As such it is commonly deployed on a hovering AUV capable of approaching man-made underwater structures at very low speeds. Feature based SLAM based on ESEIF has also been implemented with a FLS for ship hull inspections [26]. The issue of online feature extraction from FLS images is addressed in [136]. The mechanical scanning imaging sonar (MSIS) is shown in Fig. 8-d. Its operation is similar to the FLS except that instead of multiple beams, a single beam is rotated through the desired viewing angle. Consequently, the update rate is slow. It cannot be assumed that the AUV pose is constant for an entire sensor scan cycle, which increases